ATM-Bench is the first benchmark for multimodal, multi-source personalized referential memory QA over long time horizons (~4 years) with evidence-grounded retrieval and answering.
Existing long-term memory benchmarks focus primarily on dialogue history, failing to capture realistic personalized references grounded in lived experience.… See the full description on the dataset page:
https://huggingface.co/datasets/Jingbiao/ATM-Bench.
